Merge remote-tracking branch 'origin/released'

Conflicts:
	mesalib/src/mesa/drivers/dri/swrast/swrast.c
	mesalib/src/mesa/main/.gitignore
	mesalib/src/mesa/main/queryobj.c

7 files changed, 182 insertions, 588 deletions

diff --git a/mesalib/src/gallium/auxiliary/util/u_debug.c b/mesalib/src/gallium/auxiliary/util/u_debug.c
index d79f31ea9..2d2d049b2 100644
--- a/mesalib/src/gallium/auxiliary/util/u_debug.c
+++ b/mesalib/src/gallium/auxiliary/util/u_debug.c
@@ -402,31 +402,28 @@ void debug_print_format(const char *msg, unsigned fmt )
 #endif
 
 
-
-static const struct debug_named_value pipe_prim_names[] = {
-#ifdef DEBUG
-   DEBUG_NAMED_VALUE(PIPE_PRIM_POINTS),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_LINES),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_LINE_LOOP),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_LINE_STRIP),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLES),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLE_STRIP),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLE_FAN),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_QUADS),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_QUAD_STRIP),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_POLYGON),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_LINES_ADJACENCY),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_LINE_STRIP_ADJACENCY),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLES_ADJACENCY),
-   DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY),
-#endif
-   DEBUG_NAMED_VALUE_END
-};
-
-
-const char *u_prim_name( unsigned prim )
+/** Return string name of given primitive type */
+const char *
+u_prim_name(unsigned prim)
 {
-   return debug_dump_enum(pipe_prim_names, prim);
+   static const struct debug_named_value names[] = {
+      DEBUG_NAMED_VALUE(PIPE_PRIM_POINTS),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_LINES),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_LINE_LOOP),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_LINE_STRIP),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLES),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLE_STRIP),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLE_FAN),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_QUADS),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_QUAD_STRIP),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_POLYGON),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_LINES_ADJACENCY),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_LINE_STRIP_ADJACENCY),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLES_ADJACENCY),
+      DEBUG_NAMED_VALUE(PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY),
+      DEBUG_NAMED_VALUE_END
+   };
+   return debug_dump_enum(names, prim);
 }
 
 
@@ -722,38 +719,75 @@ error1:
 void
 debug_print_transfer_flags(const char *msg, unsigned usage)
 {
-#define FLAG(x)  { x, #x }
-   static const struct {
-      unsigned bit;
-      const char *name;
-   } flags[] = {
-      FLAG(PIPE_TRANSFER_READ),
-      FLAG(PIPE_TRANSFER_WRITE),
-      FLAG(PIPE_TRANSFER_MAP_DIRECTLY),
-      FLAG(PIPE_TRANSFER_DISCARD_RANGE),
-      FLAG(PIPE_TRANSFER_DONTBLOCK),
-      FLAG(PIPE_TRANSFER_UNSYNCHRONIZED),
-      FLAG(PIPE_TRANSFER_FLUSH_EXPLICIT),
-      FLAG(PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
+   static const struct debug_named_value names[] = {
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_READ),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_WRITE),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_MAP_DIRECTLY),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_DISCARD_RANGE),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_DONTBLOCK),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_UNSYNCHRONIZED),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_FLUSH_EXPLICIT),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_PERSISTENT),
+      DEBUG_NAMED_VALUE(PIPE_TRANSFER_COHERENT),
+      DEBUG_NAMED_VALUE_END
    };
-   unsigned i;
 
-   debug_printf("%s ", msg);
+   debug_printf("%s: %s\n", msg, debug_dump_flags(names, usage));
+}
 
-   for (i = 0; i < Elements(flags); i++) {
-      if (usage & flags[i].bit) {
-         debug_printf("%s", flags[i].name);
-         usage &= ~flags[i].bit;
-         if (usage) {
-            debug_printf(" | ");
-         }
-      }
-   }
 
-   debug_printf("\n");
-#undef FLAG
+/**
+ * Print PIPE_BIND_x flags with a message.
+ */
+void
+debug_print_bind_flags(const char *msg, unsigned usage)
+{
+   static const struct debug_named_value names[] = {
+      DEBUG_NAMED_VALUE(PIPE_BIND_DEPTH_STENCIL),
+      DEBUG_NAMED_VALUE(PIPE_BIND_RENDER_TARGET),
+      DEBUG_NAMED_VALUE(PIPE_BIND_BLENDABLE),
+      DEBUG_NAMED_VALUE(PIPE_BIND_SAMPLER_VIEW),
+      DEBUG_NAMED_VALUE(PIPE_BIND_VERTEX_BUFFER),
+      DEBUG_NAMED_VALUE(PIPE_BIND_INDEX_BUFFER),
+      DEBUG_NAMED_VALUE(PIPE_BIND_CONSTANT_BUFFER),
+      DEBUG_NAMED_VALUE(PIPE_BIND_DISPLAY_TARGET),
+      DEBUG_NAMED_VALUE(PIPE_BIND_TRANSFER_WRITE),
+      DEBUG_NAMED_VALUE(PIPE_BIND_TRANSFER_READ),
+      DEBUG_NAMED_VALUE(PIPE_BIND_STREAM_OUTPUT),
+      DEBUG_NAMED_VALUE(PIPE_BIND_CURSOR),
+      DEBUG_NAMED_VALUE(PIPE_BIND_CUSTOM),
+      DEBUG_NAMED_VALUE(PIPE_BIND_GLOBAL),
+      DEBUG_NAMED_VALUE(PIPE_BIND_SHADER_RESOURCE),
+      DEBUG_NAMED_VALUE(PIPE_BIND_COMPUTE_RESOURCE),
+      DEBUG_NAMED_VALUE(PIPE_BIND_COMMAND_ARGS_BUFFER),
+      DEBUG_NAMED_VALUE(PIPE_BIND_SCANOUT),
+      DEBUG_NAMED_VALUE(PIPE_BIND_SHARED),
+      DEBUG_NAMED_VALUE(PIPE_BIND_LINEAR),
+      DEBUG_NAMED_VALUE_END
+   };
+
+   debug_printf("%s: %s\n", msg, debug_dump_flags(names, usage));
 }
 
 
+/**
+ * Print PIPE_USAGE_x enum values with a message.
+ */
+void
+debug_print_usage_enum(const char *msg, unsigned usage)
+{
+   static const struct debug_named_value names[] = {
+      DEBUG_NAMED_VALUE(PIPE_USAGE_DEFAULT),
+      DEBUG_NAMED_VALUE(PIPE_USAGE_IMMUTABLE),
+      DEBUG_NAMED_VALUE(PIPE_USAGE_DYNAMIC),
+      DEBUG_NAMED_VALUE(PIPE_USAGE_STREAM),
+      DEBUG_NAMED_VALUE(PIPE_USAGE_STAGING),
+      DEBUG_NAMED_VALUE_END
+   };
+
+   debug_printf("%s: %s\n", msg, debug_dump_enum(names, usage));
+}
+
 
 #endif
diff --git a/mesalib/src/gallium/auxiliary/util/u_debug.h b/mesalib/src/gallium/auxiliary/util/u_debug.h
index 4c22fdfb6..3b2255244 100644
--- a/mesalib/src/gallium/auxiliary/util/u_debug.h
+++ b/mesalib/src/gallium/auxiliary/util/u_debug.h
@@ -467,6 +467,12 @@ void debug_dump_float_rgba_bmp(const char *filename,
 void
 debug_print_transfer_flags(const char *msg, unsigned usage);
 
+void
+debug_print_bind_flags(const char *msg, unsigned usage);
+
+void
+debug_print_usage_enum(const char *msg, unsigned usage);
+
 
 #ifdef	__cplusplus
 }
diff --git a/mesalib/src/gallium/auxiliary/util/u_format_other.c b/mesalib/src/gallium/auxiliary/util/u_format_other.c
index 85001c14f..20752d469 100644
--- a/mesalib/src/gallium/auxiliary/util/u_format_other.c
+++ b/mesalib/src/gallium/auxiliary/util/u_format_other.c
@@ -42,10 +42,7 @@ util_format_r9g9b9e5_float_unpack_rgba_float(float *dst_row, unsigned dst_stride
       float *dst = dst_row;
       const uint8_t *src = src_row;
       for(x = 0; x < width; x += 1) {
-         uint32_t value = *(const uint32_t *)src;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*(const uint32_t *)src);
          rgb9e5_to_float3(value, dst);
          dst[3] = 1; /* a */
          src += 4;
@@ -66,10 +63,7 @@ util_format_r9g9b9e5_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride
       const float *src = src_row;
       uint8_t *dst = dst_row;
       for(x = 0; x < width; x += 1) {
-         uint32_t value = float3_to_rgb9e5(src);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(float3_to_rgb9e5(src));
          *(uint32_t *)dst = value;
          src += 4;
          dst += 4;
@@ -83,10 +77,7 @@ void
 util_format_r9g9b9e5_float_fetch_rgba_float(float *dst, const uint8_t *src,
                                        unsigned i, unsigned j)
 {
-   uint32_t value = *(const uint32_t *)src;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-   value = util_bswap32(value);
-#endif
+   uint32_t value = util_cpu_to_le32(*(const uint32_t *)src);
    rgb9e5_to_float3(value, dst);
    dst[3] = 1; /* a */
 }
@@ -103,10 +94,7 @@ util_format_r9g9b9e5_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_str
       uint8_t *dst = dst_row;
       const uint8_t *src = src_row;
       for(x = 0; x < width; x += 1) {
-         uint32_t value = *(const uint32_t *)src;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*(const uint32_t *)src);
          rgb9e5_to_float3(value, p);
          dst[0] = float_to_ubyte(p[0]); /* r */
          dst[1] = float_to_ubyte(p[1]); /* g */
@@ -136,10 +124,7 @@ util_format_r9g9b9e5_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_strid
          p[0] = ubyte_to_float(src[0]);
          p[1] = ubyte_to_float(src[1]);
          p[2] = ubyte_to_float(src[2]);
-         value = float3_to_rgb9e5(p);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(float3_to_rgb9e5(p));
          *(uint32_t *)dst = value;
          src += 4;
          dst += 4;
@@ -160,10 +145,7 @@ util_format_r11g11b10_float_unpack_rgba_float(float *dst_row, unsigned dst_strid
       float *dst = dst_row;
       const uint8_t *src = src_row;
       for(x = 0; x < width; x += 1) {
-         uint32_t value = *(const uint32_t *)src;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*(const uint32_t *)src);
          r11g11b10f_to_float3(value, dst);
          dst[3] = 1; /* a */
          src += 4;
@@ -184,10 +166,7 @@ util_format_r11g11b10_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_strid
       const float *src = src_row;
       uint8_t *dst = dst_row;
       for(x = 0; x < width; x += 1) {
-         uint32_t value = float3_to_r11g11b10f(src);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(float3_to_r11g11b10f(src));
          *(uint32_t *)dst = value;
          src += 4;
          dst += 4;
@@ -201,10 +180,7 @@ void
 util_format_r11g11b10_float_fetch_rgba_float(float *dst, const uint8_t *src,
                                        unsigned i, unsigned j)
 {
-   uint32_t value = *(const uint32_t *)src;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-   value = util_bswap32(value);
-#endif
+   uint32_t value = util_cpu_to_le32(*(const uint32_t *)src);
    r11g11b10f_to_float3(value, dst);
    dst[3] = 1; /* a */
 }
@@ -221,10 +197,7 @@ util_format_r11g11b10_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_st
       uint8_t *dst = dst_row;
       const uint8_t *src = src_row;
       for(x = 0; x < width; x += 1) {
-         uint32_t value = *(const uint32_t *)src;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*(const uint32_t *)src);
          r11g11b10f_to_float3(value, p);
          dst[0] = float_to_ubyte(p[0]); /* r */
          dst[1] = float_to_ubyte(p[1]); /* g */
@@ -254,10 +227,7 @@ util_format_r11g11b10_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stri
          p[0] = ubyte_to_float(src[0]);
          p[1] = ubyte_to_float(src[1]);
          p[2] = ubyte_to_float(src[2]);
-         value = float3_to_r11g11b10f(p);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(float3_to_r11g11b10f(p));
          *(uint32_t *)dst = value;
          src += 4;
          dst += 4;
@@ -338,13 +308,9 @@ util_format_r8g8bx_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride,
       float *dst = dst_row;
       const uint16_t *src = (const uint16_t *)src_row;
       for(x = 0; x < width; x += 1) {
-         uint16_t value = *src++;
+         uint16_t value = util_cpu_to_le16(*src++);
          int16_t r, g;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
-
          r = ((int16_t)(value << 8)) >> 8;
          g = ((int16_t)(value << 0)) >> 8;
 
@@ -370,13 +336,9 @@ util_format_r8g8bx_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_strid
       uint8_t *dst = dst_row;
       const uint16_t *src = (const uint16_t *)src_row;
       for(x = 0; x < width; x += 1) {
-         uint16_t value = *src++;
+         uint16_t value = util_cpu_to_le16(*src++);
          int16_t r, g;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
-
          r = ((int16_t)(value << 8)) >> 8;
          g = ((int16_t)(value << 0)) >> 8;
 
@@ -407,11 +369,7 @@ util_format_r8g8bx_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride,
          value |= (uint16_t)(((int8_t)(CLAMP(src[0], -1, 1) * 0x7f)) & 0xff) ;
          value |= (uint16_t)((((int8_t)(CLAMP(src[1], -1, 1) * 0x7f)) & 0xff) << 8) ;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le16_to_cpu(value);
 
          src += 4;
       }
@@ -437,11 +395,7 @@ util_format_r8g8bx_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
          value |= src[0] >> 1;
          value |= (src[1] >> 1) << 8;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le16_to_cpu(value);
 
          src += 4;
       }
@@ -455,13 +409,9 @@ void
 util_format_r8g8bx_snorm_fetch_rgba_float(float *dst, const uint8_t *src,
                                      unsigned i, unsigned j)
 {
-   uint16_t value = *(const uint16_t *)src;
+   uint16_t value = util_cpu_to_le16(*(const uint16_t *)src);
    int16_t r, g;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-   value = util_bswap16(value);
-#endif
-
    r = ((int16_t)(value << 8)) >> 8;
    g = ((int16_t)(value << 0)) >> 8;
 
diff --git a/mesalib/src/gallium/auxiliary/util/u_format_yuv.c b/mesalib/src/gallium/auxiliary/util/u_format_yuv.c
index 891d99cf6..16ed8af7b 100644
--- a/mesalib/src/gallium/auxiliary/util/u_format_yuv.c
+++ b/mesalib/src/gallium/auxiliary/util/u_format_yuv.c
@@ -52,11 +52,7 @@ util_format_r8g8_b8g8_unorm_unpack_rgba_float(float *dst_row, unsigned dst_strid
       float r, g0, g1, b;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          r  = ubyte_to_float((value >>  0) & 0xff);
          g0 = ubyte_to_float((value >>  8) & 0xff);
@@ -77,11 +73,7 @@ util_format_r8g8_b8g8_unorm_unpack_rgba_float(float *dst_row, unsigned dst_strid
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          r  = ubyte_to_float((value >>  0) & 0xff);
          g0 = ubyte_to_float((value >>  8) & 0xff);
@@ -114,11 +106,7 @@ util_format_r8g8_b8g8_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_st
       uint8_t r, g0, g1, b;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          r  = (value >>  0) & 0xff;
          g0 = (value >>  8) & 0xff;
@@ -139,11 +127,7 @@ util_format_r8g8_b8g8_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_st
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          r  = (value >>  0) & 0xff;
          g0 = (value >>  8) & 0xff;
@@ -186,11 +170,7 @@ util_format_r8g8_b8g8_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_strid
          value |= float_to_ubyte(b)  << 16;
          value |= float_to_ubyte(g1) << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -206,11 +186,7 @@ util_format_r8g8_b8g8_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_strid
          value |= float_to_ubyte(b)  << 16;
          value |= float_to_ubyte(g1) << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -243,11 +219,7 @@ util_format_r8g8_b8g8_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stri
          value |= b  << 16;
          value |= g1 << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -263,11 +235,7 @@ util_format_r8g8_b8g8_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stri
          value |= b  << 16;
          value |= g1 << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -304,11 +272,7 @@ util_format_g8r8_g8b8_unorm_unpack_rgba_float(float *dst_row, unsigned dst_strid
       float r, g0, g1, b;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          g0 = ubyte_to_float((value >>  0) & 0xff);
          r  = ubyte_to_float((value >>  8) & 0xff);
@@ -329,11 +293,7 @@ util_format_g8r8_g8b8_unorm_unpack_rgba_float(float *dst_row, unsigned dst_strid
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          g0 = ubyte_to_float((value >>  0) & 0xff);
          r  = ubyte_to_float((value >>  8) & 0xff);
@@ -366,11 +326,7 @@ util_format_g8r8_g8b8_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_st
       uint8_t r, g0, g1, b;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          g0 = (value >>  0) & 0xff;
          r  = (value >>  8) & 0xff;
@@ -391,11 +347,7 @@ util_format_g8r8_g8b8_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_st
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          g0 = (value >>  0) & 0xff;
          r  = (value >>  8) & 0xff;
@@ -438,11 +390,7 @@ util_format_g8r8_g8b8_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_strid
          value |= float_to_ubyte(g1) << 16;
          value |= float_to_ubyte(b)  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -458,11 +406,7 @@ util_format_g8r8_g8b8_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_strid
          value |= float_to_ubyte(g1) << 16;
          value |= float_to_ubyte(b)  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -495,11 +439,7 @@ util_format_g8r8_g8b8_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stri
          value |= g1 << 16;
          value |= b  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -515,11 +455,7 @@ util_format_g8r8_g8b8_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stri
          value |= g1 << 16;
          value |= b  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -556,11 +492,7 @@ util_format_uyvy_unpack_rgba_float(float *dst_row, unsigned dst_stride,
       uint8_t y0, y1, u, v;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          u  = (value >>  0) & 0xff;
          y0 = (value >>  8) & 0xff;
@@ -577,11 +509,7 @@ util_format_uyvy_unpack_rgba_float(float *dst_row, unsigned dst_stride,
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          u  = (value >>  0) & 0xff;
          y0 = (value >>  8) & 0xff;
@@ -612,11 +540,7 @@ util_format_uyvy_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
       uint8_t y0, y1, u, v;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          u  = (value >>  0) & 0xff;
          y0 = (value >>  8) & 0xff;
@@ -633,11 +557,7 @@ util_format_uyvy_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          u  = (value >>  0) & 0xff;
          y0 = (value >>  8) & 0xff;
@@ -683,11 +603,7 @@ util_format_uyvy_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride,
          value |= v  << 16;
          value |= y1 << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -702,11 +618,7 @@ util_format_uyvy_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride,
          value |= v  << 16;
          value |= y1 << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -744,11 +656,7 @@ util_format_uyvy_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
          value |= v  << 16;
          value |= y1 << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -763,11 +671,7 @@ util_format_uyvy_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
          value |= v  << 16;
          value |= y1 << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -809,11 +713,7 @@ util_format_yuyv_unpack_rgba_float(float *dst_row, unsigned dst_stride,
       uint8_t y0, y1, u, v;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          y0 = (value >>  0) & 0xff;
          u  = (value >>  8) & 0xff;
@@ -830,11 +730,7 @@ util_format_yuyv_unpack_rgba_float(float *dst_row, unsigned dst_stride,
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          y0 = (value >>  0) & 0xff;
          u  = (value >>  8) & 0xff;
@@ -865,11 +761,7 @@ util_format_yuyv_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
       uint8_t y0, y1, u, v;
 
       for (x = 0; x + 1 < width; x += 2) {
-         value = *src++;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src++);
 
          y0 = (value >>  0) & 0xff;
          u  = (value >>  8) & 0xff;
@@ -886,11 +778,7 @@ util_format_yuyv_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
       }
 
       if (x < width) {
-         value = *src;
-
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         value = util_cpu_to_le32(*src);
 
          y0 = (value >>  0) & 0xff;
          u  = (value >>  8) & 0xff;
@@ -936,11 +824,7 @@ util_format_yuyv_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride,
          value |= y1 << 16;
          value |= v  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -955,11 +839,7 @@ util_format_yuyv_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride,
          value |= y1 << 16;
          value |= v  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
@@ -997,11 +877,7 @@ util_format_yuyv_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
          value |= y1 << 16;
          value |= v  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
 
          src += 8;
       }
@@ -1016,11 +892,7 @@ util_format_yuyv_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride,
          value |= y1 << 16;
          value |= v  << 24;
 
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-
-         *dst = value;
+         *dst = util_le32_to_cpu(value);
       }
 
       dst_row += dst_stride/sizeof(*dst_row);
diff --git a/mesalib/src/gallium/auxiliary/util/u_format_zs.c b/mesalib/src/gallium/auxiliary/util/u_format_zs.c
index ed45c52f9..f1ed32f1d 100644
--- a/mesalib/src/gallium/auxiliary/util/u_format_zs.c
+++ b/mesalib/src/gallium/auxiliary/util/u_format_zs.c
@@ -147,10 +147,7 @@ util_format_z16_unorm_unpack_z_float(float *dst_row, unsigned dst_stride,
       float *dst = dst_row;
       const uint16_t *src = (const uint16_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint16_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
+         uint16_t value = util_cpu_to_le16(*src++);
          *dst++ = z16_unorm_to_z32_float(value);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -170,10 +167,7 @@ util_format_z16_unorm_pack_z_float(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint16_t value;
          value = z32_float_to_z16_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_le16_to_cpu(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -190,10 +184,7 @@ util_format_z16_unorm_unpack_z_32unorm(uint32_t *dst_row, unsigned dst_stride,
       uint32_t *dst = dst_row;
       const uint16_t *src = (const uint16_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint16_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
+         uint16_t value = util_cpu_to_le16(*src++);
          *dst++ = z16_unorm_to_z32_unorm(value);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -213,10 +204,7 @@ util_format_z16_unorm_pack_z_32unorm(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint16_t value;
          value = z32_unorm_to_z16_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap16(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_le16_to_cpu(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -233,10 +221,7 @@ util_format_z32_unorm_unpack_z_float(float *dst_row, unsigned dst_stride,
       float *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z32_unorm_to_z32_float(value);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -256,10 +241,7 @@ util_format_z32_unorm_pack_z_float(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint32_t value;
          value = z32_float_to_z32_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -362,10 +344,7 @@ util_format_z24_unorm_s8_uint_unpack_z_float(float *dst_row, unsigned dst_stride
       float *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value =  util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_float(value & 0xffffff);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -383,16 +362,10 @@ util_format_z24_unorm_s8_uint_pack_z_float(uint8_t *dst_row, unsigned dst_stride
       const float *src = src_row;
       uint32_t *dst = (uint32_t *)dst_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *dst;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_le32_to_cpu(*dst);
          value &= 0xff000000;
          value |= z32_float_to_z24_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -409,10 +382,7 @@ util_format_z24_unorm_s8_uint_unpack_z_32unorm(uint32_t *dst_row, unsigned dst_s
       uint32_t *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_unorm(value & 0xffffff);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -430,16 +400,10 @@ util_format_z24_unorm_s8_uint_pack_z_32unorm(uint8_t *dst_row, unsigned dst_stri
       const uint32_t *src = src_row;
       uint32_t *dst = (uint32_t *)dst_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value= *dst;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_le32_to_cpu(*dst);
          value &= 0xff000000;
          value |= z32_unorm_to_z24_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -456,10 +420,7 @@ util_format_z24_unorm_s8_uint_unpack_s_8uint(uint8_t *dst_row, unsigned dst_stri
       uint8_t *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = value >> 24;
       }
       src_row += src_stride/sizeof(*src_row);
@@ -477,16 +438,10 @@ util_format_z24_unorm_s8_uint_pack_s_8uint(uint8_t *dst_row, unsigned dst_stride
       const uint8_t *src = src_row;
       uint32_t *dst = (uint32_t *)dst_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *dst;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_le32_to_cpu(*dst);
          value &= 0x00ffffff;
          value |= *src++ << 24;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -503,10 +458,7 @@ util_format_s8_uint_z24_unorm_unpack_z_float(float *dst_row, unsigned dst_stride
       float *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_float(value >> 8);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -524,16 +476,10 @@ util_format_s8_uint_z24_unorm_pack_z_float(uint8_t *dst_row, unsigned dst_stride
       const float *src = src_row;
       uint32_t *dst = (uint32_t *)dst_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *dst;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_le32_to_cpu(*dst);
          value &= 0x000000ff;
          value |= z32_float_to_z24_unorm(*src++) << 8;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -550,10 +496,7 @@ util_format_s8_uint_z24_unorm_unpack_z_32unorm(uint32_t *dst_row, unsigned dst_s
       uint32_t *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_unorm(value >> 8);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -571,16 +514,10 @@ util_format_s8_uint_z24_unorm_pack_z_32unorm(uint8_t *dst_row, unsigned dst_stri
       const uint32_t *src = src_row;
       uint32_t *dst = (uint32_t *)dst_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *dst;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_le32_to_cpu(*dst);
          value &= 0x000000ff;
          value |= *src++ & 0xffffff00;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -597,10 +534,7 @@ util_format_s8_uint_z24_unorm_unpack_s_8uint(uint8_t *dst_row, unsigned dst_stri
       uint8_t *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = value & 0xff;
       }
       src_row += src_stride/sizeof(*src_row);
@@ -618,16 +552,10 @@ util_format_s8_uint_z24_unorm_pack_s_8uint(uint8_t *dst_row, unsigned dst_stride
       const uint8_t *src = src_row;
       uint32_t *dst = (uint32_t *)dst_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *dst;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_le32_to_cpu(*dst);
          value &= 0xffffff00;
          value |= *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -644,10 +572,7 @@ util_format_z24x8_unorm_unpack_z_float(float *dst_row, unsigned dst_stride,
       float *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_float(value & 0xffffff);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -667,10 +592,7 @@ util_format_z24x8_unorm_pack_z_float(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint32_t value;
          value = z32_float_to_z24_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_le32_to_cpu(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -687,10 +609,7 @@ util_format_z24x8_unorm_unpack_z_32unorm(uint32_t *dst_row, unsigned dst_stride,
       uint32_t *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_unorm(value & 0xffffff);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -710,10 +629,7 @@ util_format_z24x8_unorm_pack_z_32unorm(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint32_t value;
          value = z32_unorm_to_z24_unorm(*src++);
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -730,10 +646,7 @@ util_format_x8z24_unorm_unpack_z_float(float *dst_row, unsigned dst_stride,
       float *dst = dst_row;
       const uint32_t *src = (uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_float(value >> 8);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -753,10 +666,7 @@ util_format_x8z24_unorm_pack_z_float(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint32_t value;
          value = z32_float_to_z24_unorm(*src++) << 8;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
@@ -773,10 +683,7 @@ util_format_x8z24_unorm_unpack_z_32unorm(uint32_t *dst_row, unsigned dst_stride,
       uint32_t *dst = dst_row;
       const uint32_t *src = (const uint32_t *)src_row;
       for(x = 0; x < width; ++x) {
-         uint32_t value = *src++;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
+         uint32_t value = util_cpu_to_le32(*src++);
          *dst++ = z24_unorm_to_z32_unorm(value >> 8);
       }
       src_row += src_stride/sizeof(*src_row);
@@ -796,10 +703,7 @@ util_format_x8z24_unorm_pack_z_32unorm(uint8_t *dst_row, unsigned dst_stride,
       for(x = 0; x < width; ++x) {
          uint32_t value;
          value = z32_unorm_to_z24_unorm(*src++) << 8;
-#ifdef PIPE_ARCH_BIG_ENDIAN
-         value = util_bswap32(value);
-#endif
-         *dst++ = value;
+         *dst++ = util_cpu_to_le32(value);
       }
       dst_row += dst_stride/sizeof(*dst_row);
       src_row += src_stride/sizeof(*src_row);
diff --git a/mesalib/src/gallium/auxiliary/util/u_math.h b/mesalib/src/gallium/auxiliary/util/u_math.h
index d6e83f962..8f62cac66 100644
--- a/mesalib/src/gallium/auxiliary/util/u_math.h
+++ b/mesalib/src/gallium/auxiliary/util/u_math.h
@@ -41,13 +41,7 @@
 
 #include "pipe/p_compiler.h"
 
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-
-#include <math.h>
+#include "c99_math.h"
 #include <float.h>
 #include <stdarg.h>
 
@@ -56,153 +50,15 @@ extern "C" {
 #endif
 
 
-#ifndef M_SQRT2
-#define M_SQRT2 1.41421356237309504880
-#endif
-
-
-#if defined(_MSC_VER)
-
-#if _MSC_VER < 1400 && !defined(__cplusplus)
-
-static INLINE float cosf( float f )
-{
-   return (float) cos( (double) f );
-}
-
-static INLINE float sinf( float f )
-{
-   return (float) sin( (double) f );
-}
-
-static INLINE float ceilf( float f )
-{
-   return (float) ceil( (double) f );
-}
-
-static INLINE float floorf( float f )
-{
-   return (float) floor( (double) f );
-}
-
-static INLINE float powf( float f, float g )
-{
-   return (float) pow( (double) f, (double) g );
-}
-
-static INLINE float sqrtf( float f )
-{
-   return (float) sqrt( (double) f );
-}
-
-static INLINE float fabsf( float f )
-{
-   return (float) fabs( (double) f );
-}
-
-static INLINE float logf( float f )
-{
-   return (float) log( (double) f );
-}
-
-#else
-/* Work-around an extra semi-colon in VS 2005 logf definition */
-#ifdef logf
-#undef logf
-#define logf(x) ((float)log((double)(x)))
-#endif /* logf */
-
-#if _MSC_VER < 1800
-#define isfinite(x) _finite((double)(x))
-#define isnan(x) _isnan((double)(x))
-#endif /* _MSC_VER < 1800 */
-#endif /* _MSC_VER < 1400 && !defined(__cplusplus) */
-
-#if _MSC_VER < 1800
-static INLINE double log2( double x )
-{
-   const double invln2 = 1.442695041;
-   return log( x ) * invln2;
-}
-
-static INLINE double
-round(double x)
-{
-   return x >= 0.0 ? floor(x + 0.5) : ceil(x - 0.5);
-}
-
-static INLINE float
-roundf(float x)
-{
-   return x >= 0.0f ? floorf(x + 0.5f) : ceilf(x - 0.5f);
-}
+#ifdef __cplusplus
+extern "C" {
 #endif
 
-#ifndef INFINITY
-#define INFINITY (DBL_MAX + DBL_MAX)
-#endif
 
-#ifndef NAN
-#define NAN (INFINITY - INFINITY)
+#ifndef M_SQRT2
+#define M_SQRT2 1.41421356237309504880
 #endif
 
-#endif /* _MSC_VER */
-
-
-#if __STDC_VERSION__ < 199901L && (!defined(__cplusplus) || defined(_MSC_VER))
-static INLINE long int
-lrint(double d)
-{
-   long int rounded = (long int)(d + 0.5);
-
-   if (d - floor(d) == 0.5) {
-      if (rounded % 2 != 0)
-         rounded += (d > 0) ? -1 : 1;
-   }
-
-   return rounded;
-}
-
-static INLINE long int
-lrintf(float f)
-{
-   long int rounded = (long int)(f + 0.5f);
-
-   if (f - floorf(f) == 0.5f) {
-      if (rounded % 2 != 0)
-         rounded += (f > 0) ? -1 : 1;
-   }
-
-   return rounded;
-}
-
-static INLINE long long int
-llrint(double d)
-{
-   long long int rounded = (long long int)(d + 0.5);
-
-   if (d - floor(d) == 0.5) {
-      if (rounded % 2 != 0)
-         rounded += (d > 0) ? -1 : 1;
-   }
-
-   return rounded;
-}
-
-static INLINE long long int
-llrintf(float f)
-{
-   long long int rounded = (long long int)(f + 0.5f);
-
-   if (f - floorf(f) == 0.5f) {
-      if (rounded % 2 != 0)
-         rounded += (f > 0) ? -1 : 1;
-   }
-
-   return rounded;
-}
-#endif /* C99 */
-
 #define POW2_TABLE_SIZE_LOG2 9
 #define POW2_TABLE_SIZE (1 << POW2_TABLE_SIZE_LOG2)
 #define POW2_TABLE_OFFSET (POW2_TABLE_SIZE/2)
@@ -499,7 +355,7 @@ util_half_inf_sign(int16_t x)
 #ifndef FFS_DEFINED
 #define FFS_DEFINED 1
 
-#if defined(_MSC_VER) && _MSC_VER >= 1300 && (_M_IX86 || _M_AMD64 || _M_IA64)
+#if defined(_MSC_VER) && (_M_IX86 || _M_AMD64 || _M_IA64)
 unsigned char _BitScanForward(unsigned long* Index, unsigned long Mask);
 #pragma intrinsic(_BitScanForward)
 static INLINE
@@ -541,7 +397,7 @@ unsigned ffs( unsigned u )
 static INLINE unsigned
 util_last_bit(unsigned u)
 {
-#if defined(__GNUC__)
+#if defined(HAVE___BUILTIN_CLZ)
    return u == 0 ? 0 : 32 - __builtin_clz(u);
 #else
    unsigned r = 0;
@@ -663,7 +519,7 @@ float_to_byte_tex(float f)
 static INLINE unsigned
 util_logbase2(unsigned n)
 {
-#if defined(PIPE_CC_GCC)
+#if defined(HAVE___BUILTIN_CLZ)
    return ((sizeof(unsigned) * 8 - 1) - __builtin_clz(n | 1));
 #else
    unsigned pos = 0;
@@ -683,7 +539,7 @@ util_logbase2(unsigned n)
 static INLINE unsigned
 util_next_power_of_two(unsigned x)
 {
-#if defined(PIPE_CC_GCC)
+#if defined(HAVE___BUILTIN_CLZ)
    if (x <= 1)
        return 1;
 
@@ -715,7 +571,7 @@ util_next_power_of_two(unsigned x)
 static INLINE unsigned
 util_bitcount(unsigned n)
 {
-#if defined(PIPE_CC_GCC)
+#if defined(HAVE___BUILTIN_POPCOUNT)
    return __builtin_popcount(n);
 #else
    /* K&R classic bitcount.
@@ -724,8 +580,8 @@ util_bitcount(unsigned n)
     * Requires only one iteration per set bit, instead of
     * one iteration per bit less than highest set bit.
     */
-   unsigned bits = 0;
-   for (bits; n; bits++) {
+   unsigned bits;
+   for (bits = 0; n; bits++) {
       n &= n - 1;
    }
    return bits;
@@ -784,8 +640,7 @@ util_bitreverse(unsigned n)
 static INLINE uint32_t
 util_bswap32(uint32_t n)
 {
-/* We need the gcc version checks for non-autoconf build system */
-#if defined(HAVE___BUILTIN_BSWAP32) || (defined(PIPE_CC_GCC) && (PIPE_CC_GCC_VERSION >= 403))
+#if defined(HAVE___BUILTIN_BSWAP32)
    return __builtin_bswap32(n);
 #else
    return (n >> 24) |
diff --git a/mesalib/src/gallium/auxiliary/util/u_sse.h b/mesalib/src/gallium/auxiliary/util/u_sse.h
index 642f96eb7..d4f51912a 100644
--- a/mesalib/src/gallium/auxiliary/util/u_sse.h
+++ b/mesalib/src/gallium/auxiliary/util/u_sse.h
@@ -44,33 +44,6 @@
 #include <emmintrin.h>
 
 
-/* MSVC before VC8 does not support the _mm_castxxx_yyy */
-#if defined(_MSC_VER) && _MSC_VER < 1500
-
-union __declspec(align(16)) m128_types {
-   __m128 m128;
-   __m128i m128i;
-   __m128d m128d;
-};
-
-static __inline __m128
-_mm_castsi128_ps(__m128i a)
-{
-   union m128_types u;
-   u.m128i = a;
-   return u.m128;
-}
-
-static __inline __m128i
-_mm_castps_si128(__m128 a)
-{
-   union m128_types u;
-   u.m128 = a;
-   return u.m128i;
-}
-
-#endif /* defined(_MSC_VER) && _MSC_VER < 1500 */
-
 union m128i {
    __m128i m;
    ubyte ub[16];